Searching for Interesting Plans as They Happen

1
Searching for Interesting Plans as They Happen

• Paul Cohen, Aram Galstyan
• Center for Research on Unexpected Events
• Information Sciences Institute
• University of Southern California

2
Classical Plan Recognition

• Inference problem Given a stream of observations
  and behavioral models of agents
• What are the most likely intentions of the agents
  ?
• Which plans are they pursuing ?
• At which stages of the plans are they now ?
• What will they do next ?

3
The Problem of Clutter

• Most information is useful

• Very little information is useful, but which is
  it?

4
The Hats Simulator
society in a box organizations meeting
planner unclassified large scale goal is to
minimize total costs of terrorist attacks,
information, and false arrests streaming data,
online analysis, catch them before they succeed
5
Hats in Brief

• A simulation of 105 agents (called hats). Most
  are benign, very few (e.g., 20) are known
  terrorists, some (e.g., 500) are covert
  terrorists.
• All hats activities are planned by a planner all
  plans involve meetings.
• Hats have capabilities, passed from one hat to
  another at meetings.
• Hats belong to organizations, none of which is
  known a priori
• When a task force of hats possessing capabilities
  that match a beacon's vulnerabilities meet at the
  beacon, it is destroyed.
• Terrorist plans are trees of meetings of many
  hats to pass capabilities from one to another,
  culminating in destruction of a beacon

Information broker
transaction data
Noise models
Population generator
Meeting planner
6
Hats Meeting PlannerThe Shell Game
7
Plans in Hats
The observers goal is to find this meeting tree
(target graph) before the final meeting takes
place
time
Beacon attack
8
Approach

• Generate candidate meeting graphs
• Test whether they are interesting
• Assume more accurate tests of bigger graphs
• A classic AI search problem

9
The Test Bayesian Filtering

• Track attributes over the states in candidate
  meeting trees (CMTs)
• Example did hat i acquire a capability in the
  course of a series of meetings in the CMT?
• Example has hat i become suspicious in the
  course of a sequence of meetings in the CMT?
• If so, increase the score of the CMT
• Build Bayesian Filters to track attributes across
  meetings

Candidate meeting tree
10
The Test Tracking Acquiring a Capability
Correct analysis about desired capabilities given
meetings
11
The Test Tracking the Terrorist Indicator
Variable
Analysis according to DBN on left about the state
of the terrorist indicator variable ?? given
observed meetings between 10 agents
12
The Generator Search AlgorithmsPreliminary
Experiments
13
Conclusion

• To find particular plans amidst huge numbers of
  other plans, generate and test candidate plans
• Tracking or abduction will serve as the test
• A will serve as the generator
• Try it with large-scale Hats problems
• Better generators (e.g., systematic search) and
  tests
• Empirical and mathematical analysis of error
  rates vs. effort